legends-stable-audio-3

Run Stable Audio 3 with an agent: turn a creative brief into prompts and a hardware-aware plan, generate resumable batches, and assemble verified continuous mixes. The same toolkit supports local Medium and guarded hosted Large jobs. Its operating instructions travel with the installed skill.

Agent setup (via cto-legends)

Part of the CTO Legends ecosystem. cto-legends is the only registered skill; this repo vendors a pinned copy at skills/cto-legends/SKILL.md.

Install with cto-legends install legends-stable-audio-3, then follow the module recipe the router loads. Do not register this module as its own skill.

first run · install · choose a surface · local generation · mixing · troubleshooting · all guides

legends-stable-audio-3 keeps local Medium, hosted Large, the Stable Audio web studio, the DAW plugin, and downstream mixing distinct. Its CLI handles local model access, setup, prompting, VRAM-aware planning, generation, resumable batches, guarded paid API jobs, and final audio assembly. One supported workflow can fill hours of background music by generating multiple coherent tracks and crossfading them, but the project is not limited to long-form or background music.

Stable Audio 3 Medium is not a Suno-style lyric-to-song system. The local CLI is strongest for instrumental music, music beds, samples, sound effects, adapter experiments, and long-form mixes assembled from fresh segments. The guarded hosted Large client adds text-to-audio, audio-to-audio, and inpainting. Medium can sometimes create vocal-like textures, but it does not reliably generate intelligible sung lyrics.

Why use Legends instead of approaching the model cold?

A generic request such as make a dark cinematic techno track leaves the agent to guess the surface, prompt structure, duration, BPM strategy, sampler settings, seed policy, and recovery path. Legends gives the agent a repeatable operating method:

1. Route the request to local Medium, hosted Large, the web studio, the plugin, or downstream mixing without silently substituting one for another.
2. Turn the creative brief into concrete musical metadata plus concise prose.
3. Calibrate hosted full songs at 380 seconds before expanding the paid batch; keep local hardware trials and short-cue requests distinct.
4. Choose duration, BPM, steps, CFG, and VRAM-aware segment sizes deliberately.
5. Preserve source audio, generation IDs, hashes, and resumable public-safe receipts.
6. Analyze active cues and assemble long masters without dead-air crossfades or one enormous FFmpeg graph.

That operating intelligence ships inside the skill. It is the product—not a link back to a private development vault.

Choose your surface

Goal	Legends route
Give an agent the full Stable Audio 3 operating method	Install or open the canonical skill
Run Stable Audio 3 Medium locally	Local CLI, official runtime, gated model bundle, and a hardware-appropriate PyTorch install
Use Stable Audio 3 Large through the API	Guarded large plan/generate/result workflow with explicit paid confirmation and recovery
Work in Stable Audio's web studio or DAW plugin	Bundled routing and production guidance; no false claim of end-to-end browser automation
Analyze or assemble existing tracks	Active-cue analysis and streaming crossfade mixer

Install the agent skill

Try the planner before downloading a model. The first-run guide checks the installed skill, previews a prompt, and calculates a ten-hour mix on an ordinary CPU. These steps require no GPU, model weights, or paid API account.

Source checkout

Clone the repository and open that directory in your agent client. AGENTS.md holds build and test instructions. The repo vendors the pinned cto-legends router skill at skills/cto-legends/SKILL.md; it registers no per-module skill of its own.

git clone https://github.com/avalonreset/legends-stable-audio-3.git
cd legends-stable-audio-3

To install the packaged operating bundle into a configured directory-based Agent Skills target, install the release wheel and choose the target explicitly:

python -m pip install "https://github.com/avalonreset/legends-stable-audio-3/releases/download/v0.1.2/legends_stable_audio_3-0.1.2-py3-none-any.whl"
legends-sa3 skill validate
legends-sa3 skill install --target <skills-directory>

The installer creates <skills-directory>/legends-stable-audio-3 and refuses to replace an existing copy. Client discovery locations vary, so it never guesses or edits global agent configuration.

What it does

• Helps users install the Stable Audio 3 runtime.
• Helps users download model files after they accept Stability AI's gated terms.
• Helps agents shape prompts, recipes, negative prompts, and repeatable settings.
• Detects or accepts VRAM limits and recommends practical segment lengths.
• Accepts free-form style prompts directly, with optional recipes for repeatable scaffolding.
• Previews expanded Stable Audio prompts before a long run.
• Generates one-off instrumental tracks or batches in Stable Audio 3-supported music styles.
• Manages an optional Underfit-powered LoRA Studio workflow for training custom Stable Audio 3 adapters.
• Runs local Medium prompt-to-audio batches and guarded hosted Large text-to-audio, audio-to-audio, and inpainting jobs.
• Gives agents explicit web-studio, DAW-plugin, prompting, benchmarking, and receipt guidance without pretending those are local CLI features.
• Analyzes generated source tracks for quiet heads/tails before mixing.
• Optionally renders a long continuous MP3 by streaming crossfades between tracks.
• Analyzes track heads, tails, and active cue points before long MP3 assembly.
• Keeps source tracks and manifests so interrupted runs can resume.

What it does not do

• It does not package, mirror, or redistribute Stable Audio model weights.
• It does not bypass Hugging Face gated model approval.
• It does not promise that a user's generated output is commercially safe. Users must read and comply with the model license and their local laws.
• It does not turn typed lyrics into clean sung songs. Stable Audio 3's official docs say the models do not output intelligible vocals and are not designed for speech or voice generation.
• --allow-vocals does not enable lyric singing. It only removes this wrapper's default instrumental bias so users can experiment with vocal-like textures.
• Its Apache-2.0 source license does not cover model weights, adapters, datasets, hosted services, third-party software, or generated media.

Local Medium quick start

From the source checkout on Windows PowerShell:

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[download]"
.\.venv\Scripts\legends-sa3.exe doctor
.\.venv\Scripts\legends-sa3.exe skill validate
.\.venv\Scripts\legends-sa3.exe download-model --model medium --output .\models\stable-audio-3-medium
.\.venv\Scripts\legends-sa3.exe plan --minutes 4 --vram-gb 24 --crossfade 12

On Linux or macOS:

python3 -m venv .venv
./.venv/bin/python -m pip install -e '.[download]'
./.venv/bin/legends-sa3 doctor
./.venv/bin/legends-sa3 skill validate
./.venv/bin/legends-sa3 download-model --model medium --output ./models/stable-audio-3-medium
./.venv/bin/legends-sa3 plan --minutes 4 --vram-gb 24 --crossfade 12

Before download-model, accept the Stability AI terms for the gated model and the bundled T5Gemma conditioner, then authenticate with Hugging Face yourself. This project never bundles weights or bypasses approval. Install the separate [generate] extra only after selecting the hardware-appropriate PyTorch build; local Medium is a supported CUDA path on Windows/Linux and a CPU fallback on macOS, not a claimed MPS-accelerated workflow.

Preview a hosted Stable Audio 3 Large REST request without spending credits:

legends-sa3 large plan `
  --operation text-to-audio `
  --prompt "TrackType: Music, VocalType: Instrumental, deep dub techno, 118 BPM" `
  --duration 380 --seed 42 --steps 8 --cfg-scale 1 --output-format wav

After verifying live Platform pricing and balance, load STABILITY_API_KEY from a secret store and add --confirmed-live-credits <n> --confirm-paid to a legends-sa3 large generate command. It submits the asynchronous job, polls the result endpoint, downloads and hashes the audio, and writes a secret-free JSON receipt. See the bundled references/large-api.md for text-to-audio, audio-to-audio, and inpaint examples and boundaries.

The generation ID is preserved in a pending receipt before polling. Resume an interrupted paid job with legends-sa3 large result --generation-id <id> --output <file> rather than submitting it again. Existing outputs and receipts are protected unless --overwrite is explicit, uploaded audio is preflighted with ffprobe, and downloaded bytes are validated before writing.

Preview the prompts before spending GPU time:

legends-sa3 prompt `
  --style "trance hip hop jazz, smoky saxophone, broken beat drums, 104 BPM" `
  --count 3

Generate a single prompt-guided track:

legends-sa3 generate `
  --model-dir .\models\stable-audio-3-medium `
  --stable-audio-repo ..\stable-audio-3 `
  --style "trance hip hop jazz, smoky saxophone, broken beat drums, 104 BPM" `
  --minutes 6 `
  --vram-gb 24 `
  --output .\output\trance-hip-hop-jazz

Generate a long continuous background mix by sequencing many tracks:

legends-sa3 generate `
  --model-dir .\models\stable-audio-3-medium `
  --stable-audio-repo ..\stable-audio-3 `
  --style "lo-fi study hip hop, warm Rhodes, soft boom bap drums, no vocals" `
  --hours 10 `
  --vram-gb 24 `
  --output .\output\lofi-study-10h

Generate with a native Stable Audio 3 LoRA or DoRA adapter:

legends-sa3 generate `
  --model-dir .\models\stable-audio-3-medium `
  --stable-audio-repo ..\stable-audio-3 `
  --style "digital hardcore breakbeat trance, acidic bassline, no vocals, 172 BPM" `
  --minutes 6 `
  --vram-gb 24 `
  --lora-ckpt-path .\adapters\eisbach-medium\model.safetensors `
  --lora-strength 1.0 `
  --output .\output\eisbach-breakbeat-test

Install the optional Underfit LoRA Studio bridge:

legends-sa3 lora-studio install
legends-sa3 lora-studio status

The bridge fetches only the reviewed Underfit commit 8a96800a58c0e8b82327fc04ac31c473ed900b73 and verifies the origin, commit, required files, and MIT license hash before execution. The default command only checks out code. --run-underfit-install executes downloaded third-party shell code that bootstraps/synchronizes Python tooling and can, with --with-setup, clone runtime code or download large gated model packs. Review it first and accept model terms yourself.

After training in Underfit, import a checkpoint into the Legends adapter registry:

legends-sa3 lora-studio import `
  .\state\runs\my-style\5000.safetensors `
  --name my-style `
  --source-run my-style

Mix an existing folder of MP3 tracks:

legends-sa3 mix `
  --input-dir .\output\lofi-study-10h\tracks_mp3 `
  --output .\output\lofi-study-10h\lofi-study-10h-master.mp3 `
  --crossfade 12 `
  --mix-policy active-cue

Analyze generated source tracks before mixing:

legends-sa3 analyze `
  --input-dir .\output\lofi-study-10h\tracks_mp3 `
  --json-output .\output\lofi-study-10h\track-analysis.json

Practical Defaults

The defaults are based on the RTX 4090 24 GB test run:

• Model: stabilityai/stable-audio-3-medium
• Practical single-generation length: about 380s on 24 GB VRAM
• Crossfade: 12s
• Default mix policy: active-cue
• Steps: 8
• CFG scale: 1.0
• Output: stereo MP3, 44.1 kHz, 320 kbps

For lower VRAM cards, start lower:

• 8 GB: 90s to 120s
• 12 GB: 150s to 180s
• 16 GB: 210s to 240s
• 24 GB class: 360s to 380s

Run legends-sa3 plan --hours 1 --vram-gb 16 to see the exact segment count.

Prompt-First Workflow

Stable Audio 3 responds best to concrete musical language: genre, instruments, mood or energy, BPM, and production character. legends-stable-audio-3 therefore uses free-form --style prompts as the primary workflow.

Recipes are optional scaffolds for repeatable batches. They are not a catalog of everything the model can do, and users should not need a preset before asking for something like "trance hip hop jazz" or "cinematic gospel house with breakbeats."

Current recipes:

• lofi-study
• trip-hop-trance

Vocal And Lyric Boundary

Do not position Stable Audio 3 Medium as a lyric or clean-vocal generator.

The model can be prompted with vocal language, but its own prompting guide says it does not output intelligible vocals. In practice, vocal prompts often produce vocal-like pads, chants, vowel textures, or gibberish. That can be useful for sound design, but it is not a reliable song-with-lyrics workflow.

For release examples and user-facing demos, prefer instrumental music, background beds, samples, sound effects, solo instruments, audio-to-audio edits, inpainting, continuation, and long MP3 mixes.

Long-Duration Assembly

Stable Audio 3 does not create a ten-hour file in a single generation. This operator handles that by planning multiple separate generations, keeping them in the same creative lane, and streaming them into one master with smooth crossfades. That gives users a practical way to create long background beds without looping the same song for hours.

Full-song hosted production

Hosted Large full songs default to 380 seconds. Generate one full-length candidate before expanding; preserve and reuse good recordings. Short cues and SFX use an explicit duration. The upstream API's omitted default remains 190s.

legends-sa3 production-plan --minutes 20
legends-sa3 arrange --plan arrangement.json --output arrangement-float.wav

The plan estimates four full songs, not an exact 20-minute edit. The arrangement uses individual cue points and fade lengths for each pair, records every trim, preserves full sources by default, and produces a float WAV for mastering. See full-song strategy and cue-sheet format. Automated checks do not prove creative quality; missing listening review remains explicitly unreviewed.

Legacy active-cue mixing

The mix command below retains its compatibility behavior. For complete-song sets, prefer arrange and individual musical cues instead of a uniform overlap.

Crossfade Quality Policy

The default active-cue mix policy runs per-track cue analysis before rendering:

• Trims generated near-silent dead air at source heads and tails.
• Trims quiet generated tails that would otherwise waste the overlap window.
• Keeps the first track from getting an artificial fade-up.
• Starts incoming tracks at a usable active cue so the overlap is not wasted on a long warmup.
• Leaves the final track without an artificial fade-down.
• Writes per-track trims and warnings into the MP3 manifest.

Use --mix-policy strict when you want exact raw track boundaries with only the linear overlap crossfade. Use --quality-gate fail when you want the run to stop before rendering if any track has cue-analysis warnings.

Agent support

cto-legends is the only registered skill; this repo vendors a pinned copy at skills/cto-legends/SKILL.md and registers no per-module skill. The wheel ships the operating bundle so agent installation remains available after pip install.

The installed bundle also carries its own references/ library: prompt and duration tournaments, BPM decisions, SFX and spoken-word-bed practice, exact Large REST request/polling flow, paid-action receipts, matched A/B boundaries, long-mix policy, adapter escalation, licensing boundaries, and a portable mastery eval.

The agent instructions all route users toward the same safe workflow: doctor, model access, prompt planning, generate, optional mix, verify.

After installing the Python package, use the shipped bundle without needing a repository checkout:

legends-sa3 skill validate
legends-sa3 skill install --target <skills-directory>

The installer requires an explicit parent directory and refuses to replace an existing legends-stable-audio-3 folder. It never guesses or modifies global agent locations.

Documentation

• Model access
• Prompting guide
• LoRA adapters
• LoRA Studio with Underfit
• VRAM guide
• Continuous mix workflow
• Agent compatibility
• Windows, Linux, and macOS support
• Hosted surfaces and receipts
• Stable Audio 3 Large REST reference
• Commercial use and license notes
• Apache-2.0 decision and rights record
• Dependency and reproducibility policy
• Troubleshooting
• Release checklist

Status

Current package version: v0.1.2. The project is public and the current release is available from GitHub Releases.

Project-owned source, documentation, tests, and the current banner are licensed under Apache License 2.0. NOTICE, THIRD_PARTY_NOTICES.md, and assets/PROVENANCE.md document attribution and scope. Models, weights, hosted services, datasets, adapters, generated media, and third-party components retain their separate terms.

legends-stable-audio-3 is an independent compatibility project. It is not affiliated with, sponsored by, or endorsed by Stability AI. Stability AI and Stable Audio names identify compatible services and models; their marks remain with their respective owners.

Project: cto-legends.com · Source: avalonreset/legends-stable-audio-3